I currently use what would be around half my salary at the currently listed full API prices (which I doubt we’d ever have to pay but lets assume we did), and a decent portion of that is experimentation, learning, wastage etc. I could cut it down to a third of my salary if I was more conservative. And I’m building 3x more features than I was without it. That’s a good deal even at full price, it’s just even better now with the subscription which is why the smart thing for any engineer to do right now is to take advantage of those temporary heavily subsidised prices to get as good as they can at agent accelerated coding

Never heard of a subscription model before, I take it?

Yeah on the one hand I agree with you, for example Claude Enterprise plans are going to lose the subscription option next time they annually renew, which for us isn’t for a while yet. On the other hand, the rate that new models are coming out is crazy and prices aren’t going up, the result being that every month we’re paying less and less per unit of work done. Sonnet 4.6 is as good as Opus 4.5 was but 1/3 cheaper. More and more companies outside the big 3 are becoming realistic contenders. I think if Anthropic start making pay their obviously over inflated API prices they’ll just lose a ton of customers to cheaper inference providers. Currently I use Claude Code purely because it’s the only harness you’re allowed to use on the cheap subscription plan, as soon as that plan goes away I’d jump to a different company’s plan, and if they all go away then I’d at least use a better harness (e.g. Pi) and a cheaper inference provider (e.g. DeepSeek). Meanwhile open source models are getting better and better and running our own inference farm is quite viable too

The subscription plans are insanely cheap currently, but won’t be forever. Atm you can pay $200/month to get anywhere from 1.2x to 5x faster depending where you are on the learning curve. If people aren’t taking advantage of that then it’s equivalent to if your office had a team of 10 engineers but only 5 of them are being given any tasks and the others are told to do nothing.

As to people adopting it naturally, there are any number of reasons why they wouldn’t, even if it’s good. Firstly it’s a massive change and people as a rule do not like change. Secondly it’s a learning curve, it’s not easy, and you have to invest time in it and initially it will make you slower rather than faster. Thirdly, what’s in it for the employee? They’re not gonna get paid more if they use AI to get 2x more work done, they’re employer captures all of that. They work the same hours, have to learn a whole new workflow, just go get the same pay. Why bother?

Because it’s a proxy metric for adoption. Currently you generally want your engineers to be using it as much as you can afford to so that they all learn and improve and adapt existing processes etc

Firstly why would you pay API prices when you can use the subscription plans for 20x less

Secondly that much API spend would require dozens of senior engineers using it full time, like you’d be doing the work of an entire, quite large engineering team not a single junior

Clearly the bugs weren’t actually preventing security then